Where AI actually makes money: cases from large and small companies

I have reviewed the consultancies' own cases. The money lies in the small processes.

When I asked my readers in August what they wanted more of, one answer clearly came out on top: experiences from companies. Where is AI being used, and where does it actually generate money? When I promised to collect those cases, more than ten readers wrote CASES in the comments within the first day. Nobody asked for a new model.

That matches what I see in companies. The licences have been bought, people have attended courses, and at 8 a.m. on Monday they open the same email and the same spreadsheet as the previous Friday. I call it Monday morning syndrome.

That is why I have dug out the documented cases published by the big consultancies and small Danish businesses themselves. McKinsey, BCG, PwC, Capgemini, Implement and Netcompany on one side. The case archives of the Confederation of Danish Industry (DI) and the Danish Chamber of Commerce, plus a couple of smaller Danish advisers, on the other. I have only included cases with a concrete figure, and separated measured results from promises.

The picture is clearer than I expected.

1. Half get something out of AI. Very few get a lot

First, the big numbers, so you know where the cases stand.

On September 30, BCG published a survey of 1,330 senior executives. 7.5% are what BCG calls future-built: companies where AI is changing the business. Another 41% are scaling. A year ago, only 5% were future-built. Something is happening.

McKinsey's State of AI from August traces the same curve. 37% can see an impact on operating profit, but only around 6% derive 5% or more of their EBIT from AI. That share has not changed in a year.

PwC surveyed 4,454 chief executives in January. 56% had not yet seen a significant financial benefit. And Bain surveyed more than 100 CFOs in April: 31% were satisfied with the results of their AI investments.

Adoption is broad. The gains are narrow.

Four global measurements from 2026.

Only 31% of CFOs are satisfied with the results of their AI investments.

Bain, CFO survey, April 2026.

Model 1. How far companies have come in making money from AI.

It is thought-provoking. Getting money out of AI is harder than you might think, even when the suits are lined up and expectations are sky-high. Right now, you cannot count on it reaching the bottom line by itself.

In Denmark, DI's September analysis shows that 81% of companies use AI in some form. But only 10% use it systematically in almost every department. That group grew revenue by 13.3 percent more than comparable companies from 2021 to 2025. DI is explicit that this is a correlation, not proof that AI caused it.

Key figures

81% of Danish companies use AI, but only 10% use it systematically in almost every department. That group grew revenue by 13.3 percent more than comparable companies from 2021 to 2025 (DI, September 2026; a correlation, not proof of causation).

Deloitte has a Nordic figure that should wake up a board: only a third of Nordic companies see a benefit from their best AI initiative within two years. In the rest of Europe, it is almost half.

Adoption is broad. The gains are narrow. So where are they?

2. The big consultancies find the money in customer service

Put the big consultancies' cases with measured results side by side, and the pattern is almost embarrassingly clear. Most results are in customer service and contact centres.

PwC describes the Wyndham hotel chain, where AI agents in customer service have reduced call duration by 30–50%, and AI now handles 28% of inbound calls. Reviewing hotel brand standards takes 94% less time. The person presenting the results is the director of global contact centres.

In April, McKinsey described an American credit union with around USD 10 billion in assets. Contact centre costs fell 10% in six months, and call duration 15% in 100 days. In the same article, McKinsey describes another bank where AI could safely handle only 20% of calls about car loan payments with the process as it stood. The remainder first required changes to the process and data.

BCG has the outsourcing company Konecta, operating in 28 countries. Up to 35% shorter calls and up to 40% higher sales conversion. But notice what it took: a Chief AI Deployment Officer, a new role leading hybrid teams of humans and agents, and over 25,000 employees now using the AI solutions, with an internal academy training them.

In Denmark, Netcompany built TopGPT for Topdanmark. By November 2024 it had held 80,000 customer conversations and received ten times as many positive reviews as the old chatbot. At a PwC event in June, Tryg reported around DKK 100 million less fraud annually, and around 85% of car damage liability cases now handled in minutes.

Table 1. Selected customer service cases with measured figures (self-reported).

Tryg also said something worth more than the figures: they had around 100 AI initiatives and few business cases. So they reduced them to three priority areas. That is the decision many Danish management teams need to make.

Principle

Strategy is more about what you do not choose than what you choose. Tryg went from around 100 AI initiatives to three priority areas.

3. Most chase savings. The top line is the big gap

Notice what the big consultancies' cases have in common. Most are about doing the same thing more cheaply: shorter calls, fewer calls, less fraud. Cases with measured growth in sales and revenue exist, but are clearly a minority. In BCG's own collection of AI examples from December 2024, only around one in six concerns higher sales, conversion or revenue. The remainder mostly concern costs and productivity. Growth cases are more often anonymous or described as "up to". That is striking, because BCG itself writes that 70% of the value potential lies in core functions such as sales, marketing and product development.

Most tellingly, McKinsey says so itself. In the April 2026 McKinsey Quarterly, three of its people write that AI productivity quickly becomes the entry ticket: competitors catch up with the gains, and customers receive the benefit. Lasting value comes from new offerings, business models and market structures. They end by reminding us that history is full of companies that confused efficiency with advantage. The consultancy's strategists and its cases point in different directions. That gap is an opportunity if you want to move faster than competitors.

The surveys show the same thing. At PwC, only 12% of chief executives had obtained both lower costs and higher revenue from AI. In DI's survey, 20% of companies say AI has increased their revenue. McKinsey shows what distinguishes the few who succeed. Around 80% in both groups use AI for efficiency, but 74% of high performers also target growth, compared with 47% of others. For innovation, the figures are 65% versus 49%.

Table 2. What high performers pursue. Source: McKinsey, The state of AI in 2026.

I am rather obsessed with the top line, for two reasons. First, that is where the big value lies. You realise a saving once. A better business keeps giving. The second reason is human. If an AI project effectively means tasks and jobs disappear, getting people on board is an uphill battle. Who wants to train the agent that will take over their own job? When AI is about winning more customers, making better products and giving employees new tasks, there is suddenly a reason to participate. Acceptance spreads throughout the organisation, even as agents arrive.

MIT economist and Nobel laureate Daron Acemoglu has argued the same for years. In their 2023 book Power and Progress, he and Simon Johnson warn against what Acemoglu calls so-so automation: technology that replaces humans but only makes things slightly cheaper. It creates neither much productivity nor new jobs. In February this year, Acemoglu, David Autor and Simon Johnson published Building Pro-Worker Artificial Intelligence. Their point is that AI creates most value when it makes human expertise and judgement more valuable and creates new tasks.

Replacing people with machines is the path of least resistance, and businesses risk leaving value on the table by choosing it. (Simon Johnson, paraphrased)

Examples exist, including in these cases. Konecta increased sales conversion by up to 40%. Junckers gained 66% more traffic in six months. At Hymatik, the director is clear about the goal: saved time should be used to help customers and grow, not cut staff.

My advice

Put at least one use case on the list that will move the top line, and tell employees what the freed-up time will be used for. It shifts the conversation from who becomes redundant to what we can now do together. More and more champions and enthusiasts volunteer. It is a meaningful win-win.

4. Small businesses find the money in quotations

Small Danish businesses look very different. None has a Chief AI Deployment Officer. What they do have is one painful process, and someone tackling it.

Hymatik in Odense sells hydraulics and pneumatics. Four product specialists handled quotations in Excel, and after a holiday they could be 14 days behind. Now Claude reads the customer's email, identifies the products, looks them up in the product database and retrieves prices, delivery times and data sheets. An employee approves and sends. Quotation time has halved. They have an external IT consultant two days a week, whom the director estimates costs DKK 600,000–700,000 annually and considers worth every krone.

The Vald. Birn foundry uses a model to generate an initial price proposal for incoming enquiries. A price calculation used to take up to an hour and a half per order. Now it takes almost 90% less time. The model covers 20% of enquiries; the rest are still handled manually.

At TN Værktøjsslibning, up to half of the two managers' time went on answering technical customer questions. With a closed AI connected to the ERP system, a lookup now takes perhaps 20 seconds rather than around half an hour. The solution was built in a month and a half.

Sixt Denmark hired automation company WITRICS to sort its inbox and process parking fines, among other tasks, using AI and software robots. Fines alone save over 375 hours annually. Junckers had agency No Zebra build an AI engine for landing pages with human approval: 66% more traffic in six months.

Then there are those doing it themselves. Amazon agency SUBSTANT has seven employees and has saved around half a position on weekly customer reports for a couple of hundred kroner a month. Office supplier Lomax saves 30 hours a week on product descriptions, and translating 100 products has gone from four weeks to one hour. Lomax has AI ambassadors in every department and mandatory training, and deliberately rejected chatbots in customer service.

Odense Aafart has two employees during winter. The director built booking and operational systems himself with ChatGPT and Cursor and cut licence costs by around DKK 300,000 annually.

Table 3. Danish SME cases. Sources: DI, the Danish Chamber of Commerce and WITRICS (self-reported figures).

5. Knowledge lookup is the quiet winner

There is a third category that rarely makes headlines: finding the right answer quickly.

Capgemini helped Sweden's SKF give technicians an AI that finds repair instructions. It used to take just over 12 minutes. Now it takes 90 seconds. TN Værktøjsslibning is the same story on a smaller scale.

In March, Implement built a solution for a global life science company that sorts regulatory news. 42% could be filtered out as irrelevant without losing anything important. It was tested on 1,500 historical cases, not yet measured in live operations, but it is exactly the kind of work that fills a compliance department's day.

Knowledge lookup in practice

Three cases where before-and-after time can be measured.

A bounded set of documents, a human using the answer, and a before-and-after time that can be measured.

Model 2. The common feature in knowledge lookup cases.

6. Four things recur among those who make it work

Across large and small companies, the same four things recur. They have almost nothing to do with technology.

The four common features

What recurs in the cases, large and small.

1. One process at a time. Nobody started with "AI throughout the company". Hymatik: quotations. Vald. Birn: price calculations. Sixt: parking fines.

2. Get the data right first. Without data and processes, AI cannot get far. Hymatik: years spent building a product database. McKinsey bank: only a fifth of one call type before the process was redesigned.

3. An owner in the business. Responsibility does not sit in IT. Wyndham: the contact centre director. Konecta: a senior executive responsible for AI deployment. BCG: over 80% of AI budgets sit outside IT.

4. A human approves. That is how quality is maintained, and how people dare to use it. Hymatik: an employee sends the quotation. Junckers: a human approves texts before publication.

BCG's rule of thumb: 10% algorithms · 20% technology · 70% people, organisation and processes.

7. Read the cases critically

Let me be honest about the material. Almost all figures here come from the companies or consultants themselves. None has been audited by an independent party, and many pages say nothing about how measurements were made.

Three things I look for when reading an AI case:

1. Does it say "up to"? Then it is the best figure, not the typical one. Konecta's 35% and 40% are "up to".

2. Is it measured or calculated? Haulier H. Daugaard had a model built to choose the right trailer for around 70,000 annual orders, with an estimated benefit of 4,700–5,800 hours. The model was ready but not yet integrated into the systems when the case was written, and reaching that point cost DKK 75,000 extra in consulting hours and several hundred internal hours. It is an honest case. It shows potential that has not yet been realised.

3. Who is behind it? In June, it emerged that KPMG had withdrawn a global report on agentic AI from the previous October after UBS, the Swiss Federal Railways, Transport for London and NHS Greater Manchester rejected descriptions the report gave of them. Research group GPTZero found the errors, and the Financial Times confirmed them. The month before, EY had withdrawn a report for similar reasons.

The cases are still useful. Just ask them the same questions you would ask a vendor trying to sell you something.

8. What I see in Danish companies

The companies I advise start with real cases. I support prioritisation, then implementation, then measurement points and realistic plans. Finally, we address training for leaders, teams and individual employees. In tools, I mostly see Copilot and Microsoft's Agent 365 platform as the backbone. But I insist on a hybrid setup where the customer is as independent of model and vendor as possible. I also work extensively with Anthropic's Claude, OpenAI's ChatGPT and local models for those who want full control of their data and face the strictest compliance requirements.

At an international product company, we started in finance with a workshop where leaders brought their own ideas. It produced many use cases. We scored them on six parameters: business value, technical feasibility, data maturity, regulatory risk, time to value and reuse. The rule was simple: an idea without a named business owner and funding gets closed. Today, around 25 ideas have passed that test and gained an owner. Some are in testing, and very few have reached production. That may sound modest. It is exactly the funnel shown in the cases above: narrow processes progressing one at a time.

From idea to pilot

How use cases are prioritised and measured.

Scoring: six parameters

• Business value
• Technical feasibility
• Data maturity
• Regulatory risk
• Time to value
• Reuse

Pilot: three figures agreed before building

• Value in one unit: hours, kroner or share of volume
• How much of the relevant work actually runs through the tool
• A quality threshold that must not deteriorate, such as error rate or rework

An idea without a named business owner and funding gets closed.

Model 3. The prioritisation funnel from an international product company (anonymised).

Every pilot gets the three figures agreed before anything is built. A single metric gets gamed. We also measured how AI was used before training anyone. Around one in four AI users was on a tool the company had not approved. We are now addressing that by establishing a platform that works, with guardrails, security and compliance built in. The first measurement is only a baseline. We measure again after 30 days and then quarterly.

Another customer, an international group with offices on several continents, had given every employee Copilot Chat. Before creating any training material, I interviewed six employees in five countries. Almost nobody knew where the tool was. Several instead used their private ChatGPT and Claude accounts, one for customer emails on their own phone. Monday morning syndrome had struck before training even began. The board therefore approved a four-phase programme: awareness for everyone, power users in small groups, leadership and a session for the board itself. The message is positive: you already have an approved tool, so use it.

At a public authority where I will teach a two-day course in AI agents, all ten participants have described a task from their own daily work in advance. The course is built around those tasks. Across all three customers, I build on the platform they already have, typically Microsoft, sometimes ChatGPT and Claude, but we design solutions so work can move to another model or vendor when prices or needs change. The algorithm is the 10%. The rest is ownership, measurement and habits.

I recognise the pattern in the cases. What distinguishes those who move forward is rarely the tool. It is whether someone has a mandate to change how work happens on Monday morning, and whether someone measures it.

9. Three questions for your next management meeting

If you hold the AI budget for 2027, start here:

1. Which single process costs us the most time today, and can we measure it before and after? Quotations, price calculations, customer enquiries and knowledge lookup are where the cases point.

2. Who in the business owns that process, and do they have a mandate to change it? If the answer is IT, it is the wrong name.

3. How much of the budget goes to licences, and how much goes to changing workflows and training people? If the split is the reverse of BCG's 10/20/70, you know where to start.

10. What should the next edition cover?

Cases vary greatly by industry. What works in retail rarely resembles what works in a bank, a pharmaceutical company or a municipality. Regulation, data and painful processes are different.

I can go two ways. One is more cases, sorted by industry: retail, banking and finance, manufacturing, software, pharma and life science, or the public sector (the six sectors I use in teaching at DIS Study Abroad). The other is my own method: how I work with prioritisation, measurement, change management and training so AI becomes a habit that lasts beyond Monday morning.

Write to me on LinkedIn or in the Substack comments with your industry, or write METHOD.

About the figures

My own customer examples are anonymised. All other cases and figures come from the named companies', consultancies' and organisations' own published materials. Figures are self-reported and not independently verified, and measurement methods are rarely disclosed. Amounts are reproduced in the currency used by the source. Status as of October 2026.

Sources

All sources are publicly available. Status as of October 2026.

• BCG, press release Sep. 30, 2026 (1,330 executives, 7.5% future-built): bcg.​com/​press/​30september2026-​ai-​starting-​to-​pay-​off-​companies-​generate-​value
• BCG, The Widening AI Value Gap, Oct. 2025 (5% future-built, 10/20/70): media-​publications.​bcg.​com
• BCG, Executive Perspectives: Value creation with AI, Dec. 2024 (42 quantified examples, 7 with sales/revenue growth, author's count): bcg.​com/​assets/​2024
• McKinsey, The state of AI in 2026, Aug. 2026: mckinsey.​com/​capabilities/​quantumblack/​our-​insights/​the-​state-​of-​ai
• McKinsey Quarterly, Where AI will create value, and where it won't (Montard, Diedrich and Catlin), Apr. 2026
• PwC, 29th Global CEO Survey, Jan. 2026: pwc.​com/​gx/​en/​news-​room
• Bain, CFO survey, Apr. 2026: bain.​com/​about-​bain/​media-​center
• Deloitte Denmark, Nordic AI survey, Dec. 2025: deloitte.​com/​dk
• DI, AI-virksomheder har øget både vækst og beskæftigelse, Sep. 2026: danskindustri.​dk/​arkiv/​analyser/​2026/​09
• PwC, Wyndham agentic AI case: pwc.​com/​us/​en/​library/​case-​studies/​wyndham-​agentic-​ai.​html
• McKinsey, The AI-powered bank: rewiring for excellence in customer care, Apr. 2026: mckinsey.​com/​industries/​financial-​services
• BCG, Konecta case: bcg.​com/​x/​mark-​your-​moment/​scaling-​ai-​global-​cx-​operations
• Netcompany, TopGPT for Topdanmark: netcompany.​com/​a-​new-​era-​of-​ai-​driven-​customer-​service
• Tryg, presentation at PwC event, Jun. 2026: pwc.​dk/​da/​arrangementer/​2026/​tryg.​pdf
• DI case archive, AI for alle: Hymatik, Vald. Birn, TN Værktøjsslibning, SUBSTANT, Odense Aafart, H. Daugaard, Junckers and No Zebra: danskindustri.dk/.../casearkiv-ai-for-alle
• WITRICS, Sixt case, Jun. 2026: witrics.​dk/​case
• Danish Chamber of Commerce, Lomax case, Jun. 2025: danskerhverv.​dk/​presse-​og-​nyheder
• Capgemini, SKF case, Nov. 2025: capgemini.​com/​se-​en/​story
• Implement, From regulatory noise to actionable change, Mar. 2026: implementconsultinggroup.​com/​case
• International Accounting Bulletin, KPMG drops AI report, Jun. 15, 2026: internationalaccountingbulletin.​com
• Acemoglu, Autor and Johnson, Building Pro-Worker Artificial Intelligence, The Hamilton Project, Feb. 2026: hamiltonproject.​org/​publication/​paper/​building-​pro-​worker-​ai
• MIT Sloan, Pro-worker AI, explained, Jun. 2026: mitsloan.​mit.​edu/​ideas-​made-​to-​matter/​pro-​worker-​ai-​explained
• Acemoglu and Johnson, Power and Progress, PublicAffairs 2023
• McKinsey, The state of AI in 2026, exhibit on objectives: efficiency, growth and innovation (high performers versus others)

First published on Substack October 18, 2026.